WSIS Forum 2026
Rapport généré par l'IA

From Data to Implementation: Scaling Digital Tools for Human Rights Monitoring

7 intervenants
Résumé

Résumé

Cette discussion, modérée par Domenico Zipoli du Geneva Human Rights Hub, portait sur la manière dont les outils numériques et l'intelligence artificielle peuvent améliorer le suivi et la mise en œuvre des droits de l'homme, en réunissant des perspectives issues des gouvernements, des institutions internationales et du monde académique . Roberto Cespedes, Chargé d'Affaires de la mission du Costa Rica auprès de l'ONU à Genève, a décrit l'expérience de son pays avec la Base de données nationale de suivi des recommandations (NRTD), introduite pour aider à gérer le volume important de recommandations reçues des organes de traités, de l'EPU et des procédures spéciales . Il a souligné que la fonctionnalité de regroupement de l'outil a amélioré la communication interinstitutionnelle et que le Costa Rica s'emploie à donner à la société civile un accès direct à la base de données afin de renforcer la transparence et la mise en œuvre collaborative . Marie Eve Boyer du HCDH a expliqué que la NRTD a été développée en réponse à la nature fragmentée des recommandations internationales relatives aux droits de l'homme, en s'appuyant sur l'Index universel des droits de l'homme pour créer un outil permettant aux États de regrouper les recommandations, d'en attribuer la responsabilité et d'en rendre compte . Elle a souligné que l'IA peut aider à identifier les données pertinentes et à élargir la portée des rapports, mais a insisté sur le fait que l'expertise humaine demeure essentielle, l'IA ne pouvant remplacer le jugement contextuel nécessaire pour évaluer la mise en œuvre concrète . À ce jour, 20 pays utilisent activement la NRTD, et 40 autres attendent son déploiement . Lukasz Szoszkiewicz a présenté une perspective académique, en démontrant des outils qu'il a développés à l'aide du traitement automatique du langage naturel et de l'IA générative pour améliorer l'accès aux bases de données des droits de l'homme de l'ONU, notamment la recherche au niveau des paragraphes et des tableaux de bord analytiques . Il a soutenu que l'IA générative a transformé le développement logiciel, permettant désormais aux experts du domaine, plutôt qu'aux professionnels de l'informatique, de diriger la conception et la mise en œuvre d'outils spécialisés . La discussion s'est conclue avec les participants reconnaissant que l'accessibilité des données et le défi de mesurer les résultats concrets demeurent des lacunes importantes, les intervenants s'accordant sur le fait que la collaboration entre les gouvernements, la société civile et le monde académique est essentielle pour rendre la mise en œuvre des droits de l'homme plus efficace, inclusive et responsable .

Points clés
  • Points clés

  • Objectif général

  • La discussion vise à explorer comment les outils numériques et l'intelligence artificielle peuvent être utilisés pour améliorer le suivi et la mise en œuvre des droits de l'homme. Elle examine plus précisément l'application pratique des bases de données de suivi, du traitement des données assisté par IA et des expérimentations académiques avec le traitement automatique du langage naturel, dans le but de rendre les recommandations relatives aux droits de l'homme plus accessibles, plus exploitables et plus efficacement mises en œuvre par les gouvernements, la société civile et les organes internationaux.
  • --
  • Principaux points de discussion

  • Le défi de la gestion de grands volumes de recommandations relatives aux droits de l'homme et le rôle des outils numériques de suivi pour y répondre. Les États reçoivent chaque année des milliers de recommandations émanant de multiples organes - organes de traités, EPU, procédures spéciales et mécanismes régionaux - ce qui rend la coordination extrêmement difficile. La Base de données nationale de suivi des recommandations (NRTD), développée par le HCDH, a été conçue spécifiquement pour consolider ces recommandations en un seul endroit, permettant aux gouvernements de les regrouper par thème, d'attribuer les responsabilités institutionnelles et de suivre les progrès de la mise en œuvre.
  • L'expérience pratique du Costa Rica en tant qu'utilisateur gouvernemental de la NRTD, mettant en lumière à la fois les avantages et les défis persistants. Le Costa Rica a créé une commission interinstitutionnelle en 2011 pour coordonner le suivi des recommandations, mais manquait d'outils techniques efficaces jusqu'à l'adoption récente de la NRTD. L'outil a amélioré la visibilité des recommandations entre les ministères et facilité la communication interinstitutionnelle grâce au regroupement thématique. Un défi persistant est le fort taux de rotation du personnel, auquel le gouvernement répond par des cycles de formation continue. Le Costa Rica s'emploie également à donner à la société civile un accès direct à la NRTD afin d'améliorer la transparence et la mise en œuvre collaborative. - Le rôle de l'IA dans l'élargissement du suivi des droits de l'homme, équilibré par le besoin irremplaçable du jugement humain et de l'expertise contextuelle. L'IA et l'apprentissage automatique peuvent traiter de grands volumes de données, identifier des tendances, soutenir la recherche sémantique et faciliter l'accès multilingue. Cependant, les intervenants ont constamment souligné que l'IA devrait soutenir plutôt que remplacer le jugement humain, et que le contexte, la nuance juridique et les voix des communautés concernées restent essentiels. La position de l'ONU, telle qu'exprimée par le HCDH, est que les humains doivent conserver le contrôle, l'IA servant d'outil d'assistance plutôt que de décideur. - Les expérimentations académiques avec le traitement automatique du langage naturel et l'IA générative pour construire des bases de données pratiques sur les droits de l'homme et des outils analytiques. Lukasz Szoszkiewicz a décrit une série d'outils prototypes construits à l'aide de l'IA générative, notamment des bases de données de jurisprudence des organes de traités de l'ONU, un tableau de bord pour l'analyse des données de l'Index universel des droits de l'homme, et un outil de recherche au niveau des paragraphes pour la Cour européenne des droits de l'homme. Une observation clé est que l'IA générative permet aux experts du domaine - tels que les juristes et les professionnels des droits de l'homme - de développer des logiciels sur mesure sans dépendre fortement des équipes informatiques, déplaçant ainsi l'équilibre du développement logiciel vers ceux qui possèdent une expertise substantielle. Il a également mis en évidence des stratégies pour réduire les hallucinations de l'IA, telles que l'utilisation de grands modèles de langage pour générer des scripts déterministes plutôt que de les interroger directement avec des données.
  • L'accessibilité des données, l'interopérabilité et le fossé entre la disponibilité des données et la mesure significative des résultats. Plusieurs intervenants ont noté que si les données existent souvent, elles sont fragmentées, pas toujours interopérables et difficiles à traiter à grande échelle. Une préoccupation particulière soulevée était la difficulté de mesurer les résultats - par opposition aux engagements ou aux processus - sans données fiables, désagrégées et au niveau communautaire, notamment dans des domaines sensibles tels que la torture. Le HCDH mène actuellement des recherches sur les ensembles de données minimaux nécessaires pour suivre les progrès, en mettant l'accent sur le lien entre les indicateurs des droits de l'homme et les indicateurs de résultats des ODD. ---
  • Ton général

  • Le ton général de la discussion est constructif, collaboratif et prudemment optimiste. Les intervenants partagent un enthousiasme sincère pour le potentiel des outils numériques et de l'IA à améliorer la mise en œuvre des droits de l'homme, tout en reconnaissant constamment les limites du monde réel telles que les lacunes dans les données, la fragmentation, les contraintes de capacité et le risque d'hallucination de l'IA. Le ton est resté pratique et ancré dans la réalité, notamment lorsque Roberto Cespedes décrit l'expérience vécue du Costa Rica , et lorsque Marie Eve Boyer souligne le rôle indispensable de l'expertise humaine aux côtés de la technologie. Vers la fin de la session, lorsque les membres du public soulèvent des défis plus complexes - tels que l'accessibilité des données pour le suivi de la torture et la mesure des résultats - le ton devient légèrement plus sobre et réflexif, tout en restant orienté vers les solutions et tourné vers l'avenir. Tout au long de la discussion, on perçoit un fort esprit de dialogue intersectoriel entre gouvernements, organisations internationales et monde académique, les intervenants invitant ouvertement à une collaboration accrue.
Intervenants
DZ
Domenico Zipoli
145 wpm · 14 min
ME
Marie Eve Boyer
166 wpm · 11 min
RC
Roberto Cespedes
123 wpm · 9 min
LS
Lukasz Szoszkiewicz
171 wpm · 12 min
AL
Axel Leblois
155 wpm · 2 min
CD
Cecilia de Armas
182 wpm · 2 min
ML
Michaela Lissowsky
124 wpm · 1 min

Résumé élargi : Des données à la mise en œuvre - Mise à l'échelle des outils numériques pour le suivi des droits humains

#

Ouverture et cadrage

La session s'est ouverte sur de brèves remarques de Michaela Lissowsky, de la Friedrich Naumann Foundation, qui a observé que l'intelligence artificielle est actuellement dominée par un petit nombre d'acteurs technologiques puissants, et a soutenu que cela rend d'autant plus important de placer les droits humains - qui appartiennent à toutes les personnes dans le monde - au cœur de l'innovation numérique . Elle a présenté la discussion comme la continuation d'une collaboration avec le Haut-Commissariat des Nations Unies aux droits de l'homme (HCDH) et le Geneva Human Rights Hub, en soulignant que, tandis que la session de l'année précédente avait introduit les outils numériques et leurs implications, cette session se concentrerait sur leur application concrète .

Domenico Zipoli, Responsable des programmes au Geneva Human Rights Hub et modérateur de la session, a exposé l'idée directrice avec clarté : la mise en œuvre des droits humains dépend de plus en plus non seulement de la volonté politique, mais aussi de la qualité des systèmes d'information qui la soutiennent . Il a noté que les États reçoivent des milliers de recommandations par an émanant des mécanismes onusiens des droits de l'homme, des organes régionaux et d'autres processus, qui doivent toutes être regroupées, attribuées, suivies et traduites en actions concrètes . Ce volume de recommandations, a-t-il soutenu, est précisément là où les solutions numériques sont devenues de plus en plus importantes, contribuant à réduire les doublons, à soutenir l'établissement de rapports, à améliorer l'accès aux recommandations et à se rapprocher de leur mise en œuvre .

Domenico Zipoli a également reconnu les limites du paysage actuel. L'écosystème des outils numériques reste fragmenté, avec de nombreux outils qui fonctionnent en silos, des données qui ne sont pas toujours interopérables, et un accès inégal pour la société civile et les défenseurs des droits humains - en particulier en dehors de l'Europe . Il s'est ensuite penché sur le rôle de l'intelligence artificielle, notant que l'IA et l'apprentissage automatique peuvent aider à traiter de grands volumes d'informations, à identifier des tendances, à regrouper des recommandations, à soutenir la recherche sémantique, à faciliter l'accès multilingue et à aider les utilisateurs à identifier des lacunes . Cependant, il a pris soin de formuler la question centrale de la session non pas comme celle de savoir si l'IA peut rendre le suivi des droits humains plus rapide, mais plutôt comment les outils numériques et l'IA responsable peuvent rendre la mise en œuvre des droits humains plus efficace, inclusive, transparente et responsable . Ce cadrage normatif - ancrant la discussion dans les valeurs des droits humains plutôt que dans les capacités techniques - a donné le ton à tout ce qui a suivi.

---

#

La perspective gouvernementale : l'expérience du Costa Rica avec la NRTD

Roberto Cespedes, Chargé d'affaires de la Mission du Costa Rica auprès des Nations Unies à Genève, a offert un compte rendu détaillé du parcours de son pays dans la gestion des recommandations relatives aux droits humains. Le Costa Rica a créé une commission interinstitutionnelle en 2011 - le Mécanisme national de suivi (NMIRF) - coordonnée par le Ministère des affaires étrangères et réunissant la plupart des ministères gouvernementaux, des représentants du pouvoir judiciaire et du parlement, ainsi que l'institution nationale des droits de l'homme (la Defensoría de los Habitantes) . Plus récemment, la commission a institutionnalisé la participation de deux représentants de la société civile par le biais de ce que l'on appelle l'Entité permanente de consultation, leur accordant une place à chaque réunion .

Malgré cette structure institutionnelle, Roberto Cespedes a reconnu que le Costa Rica avait manqué d'outils techniques efficaces pendant de nombreuses années . L'introduction de la base de données nationale de suivi des recommandations (NRTD), développée avec le soutien du HCDH, a considérablement amélioré la situation . L'outil a donné une plus grande visibilité aux recommandations des organes de traités, de l'Examen périodique universel (EPU) et des procédures spéciales, et a aidé les institutions à comprendre où elles ont des possibilités de mettre en œuvre les recommandations . Roberto Cespedes a décrit la NRTD comme flexible et intuitive, notant que des ministères, notamment ceux de l'Éducation, de la Santé et l'institution chargée de la protection de l'enfance, ont trouvé ses fonctionnalités utiles .

L'un des avantages pratiques les plus significatifs identifiés par Roberto Cespedes est la fonctionnalité de regroupement de l'outil, qui classe les recommandations par sujet ou thème. Cela a conduit les entités à reconnaître que certaines recommandations créent des opportunités de collaboration avec d'autres institutions, favorisant ainsi la communication interministérielle et, en définitive, l'action . Il a également mis en évidence un défi structurel persistant : le fort taux de rotation du personnel dans les institutions gouvernementales signifie que les connaissances institutionnelles sur les recommandations sont fréquemment perdues lors des changements de personnel . La réponse du Costa Rica a été d'instaurer un cycle de formation continue pour le personnel de tous les ministères, intégrant la NRTD dans la pratique institutionnelle courante plutôt que de la traiter comme un déploiement ponctuel .

Pour l'avenir, Roberto Cespedes s'est montré optimiste quant au potentiel de la détection de tendances assistée par l'IA au sein de la NRTD pour révéler comment différentes institutions interprètent et répondent aux recommandations . Il a également souligné que la société civile est un élément central de la stratégie de mise en œuvre du Costa Rica, et que le pays travaille activement à donner aux organisations de la société civile un accès direct à la NRT, non pas simplement en tant qu'observatrices, mais en tant que partenaires actives pouvant voir ce que chaque partie de l'État fait, formuler des suggestions et travailler en collaboration avec le gouvernement sur la mise en œuvre .

---

#

La perspective du développeur : le HCDH et la conception de la NRTD

Marie Eve Boyer, Chargée des droits de l'homme au sein du Programme de renforcement des capacités du HCDH, a exposé à la fois la justification institutionnelle et la philosophie de conception de la NRTD. Elle a commencé par décrire le problème pour lequel l'outil a été créé : les États font face à un ensemble écrasant et fragmenté de recommandations émanant simultanément de multiples organes internationaux et régionaux . Prenant le Brésil comme illustration, elle a noté qu'un seul pays peut recevoir des recommandations de la Commission interaméricaine, de la Cour interaméricaine, du Conseil des droits de l'homme, de l'EPU, des comités des organes de traités et de plusieurs rapporteurs spéciaux - cinq rapporteurs spéciaux ayant visité le Brésil au cours d'une seule année . Le défi n'est pas seulement lié au volume, mais aussi à la cohérence : d'un point de vue humain, le système peut paraître profondément fragmenté, même si le regroupement des recommandations révèle souvent qu'elles pointent dans la même direction avec peu de véritables contradictions .

La réponse du HCDH a commencé avec l'Index universel des droits de l'homme (IUDH), une plateforme en ligne qui consolide tous les résultats du système onusien, consultable par pays, thème, groupe et Objectif de développement durable (ODD) . Cependant, Marie Eve Boyer a expliqué que les États avaient besoin de plus qu'un outil de référence - ils avaient besoin d'un mécanisme pour passer de la définition de normes et des orientations à la mise en œuvre concrète . C'est ce qui a conduit au développement de la NRTD, qui extrait les recommandations relatives à un pays spécifique de l'IUDH et fournit des fonctionnalités permettant de les regrouper par thème, ODD, groupe et ministère, d'attribuer les responsabilités institutionnelles et de rendre compte des progrès accomplis .

Marie Eve Boyer a souligné que la responsabilisation est intégrée dans la conception de l'outil : en exigeant l'identification du ministère chef de file, du co-chef de file et de celui qui soutient la mise en œuvre de chaque recommandation, la NRTD rend les responsabilités explicites et traçables . Actuellement, 20 pays utilisent activement la NRTD, et environ 40 autres attendent son déploiement - une indication claire de la demande, mais aussi des contraintes de capacité auxquelles le HCDH est confronté pour accompagner les États dans ce processus .

Sur la question de l'IA, Marie Eve Boyer a clairement exprimé la position institutionnelle de l'ONU : les humains doivent conserver le contrôle, l'IA servant d'outil d'assistance plutôt que de décideur . L'IA peut aider à intensifier le travail, à repérer les données pertinentes, à identifier les données nécessaires pour mesurer les progrès et à signaler les données supplémentaires à collecter . Cependant, elle a insisté sur le fait que la composante humaine derrière les outils numériques n'est pas seulement souhaitable, mais essentielle. L'IA ne peut pas créer la réalité qui affecte les gens au quotidien, et des personnes travaillant avec des communautés qui comprennent le sujet sont nécessaires non seulement pour saisir des données, mais aussi pour les analyser et réfléchir à ce qui fonctionne dans un contexte national donné . Elle a cité des exemples d'outils qui avaient montré de réelles limites précisément parce qu'il n'y avait pas de composante humaine soutenue derrière eux pour alimenter le système en informations - un point qu'elle a formulé en référence à la reconnaissance implicite antérieure de Roberto Cespedes des lacunes d'autres outils . Cela l'a amenée à s'opposer - avec diplomatie mais clarté - à la suggestion que les outils numériques peuvent se substituer aux structures institutionnelles ou les précéder, en soutenant que le HCDH ne croit pas que cette approche fonctionne à long terme .

---

#

La perspective académique : traitement du langage naturel et IA générative en pratique

Lukasz Szoszkiewicz, Professeur assistant à l'Université Adam Mickiewicz de Poznań, Directeur des affaires européennes de la NeuroRise Foundation, et également affilié à Heuridox - un partenaire proche du Geneva Human Rights Hub - a apporté une perspective académique expérimentale à la discussion . Il s'est décrit comme un expérimentateur dont le rôle est d'explorer de nouveaux territoires et de transmettre les résultats à d'autres parties prenantes, où les risques peuvent être correctement évalués avant que les outils ne soient déployés à grande échelle . Il a souligné qu'il est juriste de formation, et non informaticien, et qu'il a appris à utiliser ces outils par l'expérimentation autodidacte - un point qu'il a délibérément soulevé pour illustrer que les experts du domaine peuvent et doivent s'engager directement dans le développement d'outils d'IA . Il a présenté quatre outils prototypes qu'il a développés en utilisant le traitement du langage naturel et l'IA générative, notamment une base de données de commentaires généraux et de jurisprudence des organes de traités de l'ONU, un tableau de bord pour l'analyse des données de l'Index universel des droits de l'homme, et un outil de recherche au niveau des paragraphes pour la Cour européenne des droits de l'homme .

Une idée centrale de la présentation de Lukasz Szoszkiewicz est que l'IA générative a fondamentalement transformé le processus de développement de logiciels. Auparavant, les experts du domaine devaient traduire leurs besoins en spécifications techniques et les confier à des équipes informatiques, attendant - parfois à un coût considérable - les outils qui en résultaient . Désormais, l'IA générative permet aux experts du domaine de construire directement des prototypes fonctionnels, de tester des interfaces utilisateur et d'aboutir à des solutions qui répondent exactement à leurs besoins . Il a illustré ce point avec une diapositive visuelle représentant l'évolution du développement de logiciels, utilisant un diagramme en couches pour contraster les approches traditionnelles et assistées par l'IA. Il a soutenu de manière cruciale que les professionnels de l'informatique sans expertise en droits humains ne peuvent pas déterminer quelles fonctionnalités sont nécessaires - par exemple, qu'une recherche au niveau des paragraphes plutôt qu'au niveau des documents est essentielle pour la recherche juridique - et que l'espace pour les experts du domaine dans le développement de logiciels s'élargit donc tandis que le rôle relatif de l'informatique diminue . L'élaboration de critères d'évaluation pour déterminer si les outils d'IA fonctionnent comme prévu est également une tâche qui ne peut être accomplie que par des personnes ayant des connaissances en droits humains . Il a également noté que les étudiants en droit, contrairement aux professionnels de l'informatique, sont souvent découragés lorsque les outils échouent ou produisent des erreurs, et que changer cet état d'esprit est une partie importante de la capacitation des experts du domaine à s'engager avec les outils d'IA.

Lukasz Szoszkiewicz a également abordé la préoccupation critique concernant les hallucinations de l'IA avec une précision technique. Il a établi une distinction essentielle entre l'utilisation de grands modèles de langage (LLM) pour analyser directement des données - ce qui risque de provoquer des hallucinations - et l'utilisation de LLM pour générer des scripts de code déterministes qui traitent les données de manière fiable et cohérente . Dans cette dernière approche, le code généré peut être inspecté par des professionnels de l'informatique pour détecter des problèmes de cybersécurité et de confidentialité avant le déploiement, et produira le même résultat à chaque fois, éliminant la variabilité probabiliste qui cause les hallucinations . Lorsqu'il utilise des LLM pour la similarité sémantique dans ses outils, il invoque les paragraphes sources mot pour mot plutôt que de générer un nouveau texte, réduisant le risque d'hallucination à un niveau très faible . Il a reconnu qu'éliminer entièrement les hallucinations reste difficile et que le processus est long et délicat, mais a maintenu que des choix de conception soigneux peuvent rendre le risque gérable .

Lukasz Szoszkiewicz a conclu avec une réflexion sur son approche : il a comparé le processus d'exploration des capacités de l'IA à la navigation dans un jeu de rôle (RPG) où l'on ne peut pas voir la carte complète, seulement son environnement immédiat, et où l'on doit explorer par essais et erreurs . Il a noté que l'un des outils accessibles aux participants avait été construit la veille au soir, en environ deux à trois heures - une illustration frappante de l'accessibilité qu'est devenu le développement d'outils sophistiqués pour les experts du domaine prêts à expérimenter.

---

#

Thèmes transversaux : conception, institutionnalisation et l'avenir des rapports

Plusieurs thèmes transversaux importants ont émergé des échanges entre les intervenants. Domenico Zipoli a observé que les outils numériques ne sont pas des contenants neutres, et que les choix de conception effectués dès le départ influenceront la manière dont les recommandations sont mises en œuvre dans la pratique - rendant une conception soigneuse et délibérée essentielle dès le début . Il a également formulé une provocation spéculative - reconnaissant qu'il exagère peut-être - selon laquelle le concept de rapports périodiques, pierre angulaire du système international des droits humains, pourrait dans cinq à dix ans céder la place à des rapports continus, à mesure que les outils numériques rendent l'information accessible de manière permanente et ouverte . Cette spéculation a relié la discussion technique sur les flux de données et l'interopérabilité à une question bien plus large sur l'architecture future de la responsabilisation internationale en matière de droits humains.

Domenico Zipoli a également introduit une observation inattendue sur la relation entre la numérisation et la construction institutionnelle : l'une des voies vers l'établissement d'un Mécanisme national de mise en œuvre, de rapport et de suivi (NMRF) consiste à numériser le travail de suivi par le biais d'outils, car l'interface numérique oblige les gouvernements à identifier les principaux acteurs de mise en œuvre, institutionnalisant ainsi des responsabilités qui pourraient autrement rester informelles . Cette observation a créé une tension productive avec la mise en garde de Marie Eve Boyer selon laquelle les outils sans soutien institutionnel humain préexistant ont historiquement échoué - un désaccord qui, bien que non entièrement résolu, a enrichi la discussion en introduisant une complexité autour du séquençage des outils et des institutions.

---

#

Contributions du public : lacunes en matière de données, mesure des résultats et société civile

Le segment de questions-réponses a introduit deux perspectives importantes depuis la salle. Cecilia de Armas, représentant le Global Torture Index à l'OMCP, a décrit les défis liés à la collecte de données dans un contexte où les gouvernements peuvent délibérément dissimuler des informations sur la torture . Elle a noté que le Global Torture Index travaille avec 100 ONG dans 39 pays, avec des plans d'expansion, mais fait face au double défi de traiter de grands volumes de données et de communiquer les résultats de manière significative au grand public . Elle a appelé davantage de gouvernements à publier des données sur les recommandations relatives à la torture, et a souligné que le dialogue multipartite entre gouvernements, société civile et monde académique est la seule voie viable pour suivre les progrès - tout en exprimant le souhait d'éviter la désignation et la stigmatisation et de mettre en lumière les évolutions positives .

Axel Leblois de G-State, fort de 20 ans de suivi de la Convention relative aux droits des personnes handicapées (CDPH) dans 143 pays, a introduit le défi méthodologique le plus rigoureux de la session . Il a noté que si l'IA peut désormais recueillir des données sur les engagements et les processus de manière relativement fiable, le domaine des résultats - reflétant les expériences réelles des personnes censées être protégées - présente un défi majeur, car sans retours d'utilisateurs validés et analyses documentées des utilisateurs finaux, l'IA entre dans un vaste champ d'hallucinations . Il a également mis en évidence un désalignement structurel entre les longs cycles d'examen des organes de traités de l'ONU et le rythme rapide des changements sur le terrain , remettant implicitement en question l'adéquation de l'architecture de suivi actuelle pour la protection des droits humains en temps réel.

---

#

Réflexions finales : données, résultats et la voie à suivre

Dans leurs réponses finales, les trois panélistes ont abordé ces défis avec un mélange de franchise et d'optimisme prudent. Lukasz Szoszkiewicz a réitéré que la disponibilité des données est le principal goulot d'étranglement : une fois que les données existent, l'IA peut les structurer et les traiter efficacement, mais sans données, le système ne peut pas fonctionner . Il a maintenu que les hallucinations peuvent être réduites à des niveaux très faibles grâce à une conception soigneuse, bien qu'il ait reconnu que cela reste un processus long et délicat .

Marie Eve Boyer a décrit les recherches en cours du HCDH sur les ensembles de données minimaux requis par les organes de traités, en commençant par les recommandations du Comité contre la torture, comme une étape pratique vers la fourniture aux États d'orientations sur la collecte de données prioritaires . Elle a suggéré que l'exploitation des indicateurs de résultats des ODD pourrait aider à établir un ensemble de données minimum de référence, offrant une voie concrète pour aligner le suivi des droits humains avec l'infrastructure de données existante .

Roberto Cespedes a formulé une observation particulièrement stimulante : de nombreuses actions gouvernementales positives ne sont pas signalées parce que les fonctionnaires ne reconnaissent pas que ce qu'ils font correspond à une recommandation relative aux droits humains . Il a proposé que l'IA pourrait aider à identifier une telle conformité cachée en analysant les données saisies et en signalant si une action gouvernementale correspond à une recommandation . Il a également suggéré que le suivi continu de la satisfaction des communautés et son lien avec les actions gouvernementales sur le terrain pourraient devenir plus faciles et moins coûteux dans un avenir proche, offrant une direction prometteuse pour la mesure des résultats .

---

#

Conclusion

La session a démontré un haut degré de consensus entre les intervenants représentant le gouvernement, les organisations internationales, le monde académique et la société civile sur les principes fondamentaux régissant l'utilisation responsable des outils numériques et de l'IA dans le suivi des droits humains. Tous les intervenants ont convenu que les outils numériques sont essentiels pour gérer le volume écrasant de recommandations émanant de multiples organes ; que l'IA doit soutenir plutôt que remplacer le jugement humain ; que l'expertise humaine et les structures institutionnelles sont irremplaçables ; et que l'accessibilité des données reste le principal goulot d'étranglement . Des défis non résolus importants subsistent, notamment la manière d'obtenir des données valides sur les résultats au niveau communautaire, comment remédier au désalignement entre les cycles d'examen des organes de traités et les conditions en temps réel, comment assurer l'interopérabilité entre des écosystèmes d'outils fragmentés, et comment étendre le déploiement aux nombreux pays en attente de soutien . La session s'est clôturée par une invitation à poursuivre la conversation au stand numéro 117 à l'Expo AI for Good, reflétant l'esprit général de collaboration intersectorielle qui a caractérisé l'ensemble de la discussion .

Michaela Lissowsky
Right now, AI is dominated by a couple of tech bros, which was one of the learning lessons I would say repeated again and again within the last days. But human rights belong to all people worldwide. Therefore, we need today's discussion about from data to implementation, scaling digital tools for human rights monitoring. I'm happy that we continue here our cooperation from last year with the Office of the High Commissioner and especially our friends and partners from the Geneva Human Rights Hub. While we introduced last year the digital tools and discussed its implications, today's next 45 minutes will be about its real -world application. And therefore, I'm happy that we have a special... Actually, you're also a member state representative amongst these groups. And me, as the German here from the Friedrich Naumann Foundation, I will take care of the time. And in order to fulfill the clichés, I will hand over to Domenico, our human rights friend from the Geneva Human Rights Hub, who will be the
Domenico Zipoli
moderator. Thank you. Thank you very much, Michaela, for these opening remarks. Thank you all for being here. Good morning to everybody in the room and online. It's a great pleasure to have this conversation about this trajectory between data and implementation, and really looking at how scaling digital tools can benefit human rights monitoring. As Michaela mentioned, my name is Domenico Zipoli. I'm head of programs of the Geneva Human Rights Hub. And I'll have the pleasure of speaking to you today. Thank you. We have 45 minutes together. The infamous Wyss is 45 minutes. I'll just use five minutes just to give a little bit of an introduction to the session. But yes, the format will be quite focused. So after my framing remarks, I will turn to our three intervenants of the day, and each speaker will have around five minutes, as mentioned, back to back, and then we will open the floor for questions and discussions with the room as well as with our colleagues online. I already have seen that there's quite a number of interesting initiatives amongst our friends in the room. It would be great to also learn from what you're doing in the digital human rights space. The guiding idea for today is quite simple. Human rights implementation increasingly depends not only on political will, but also on the quality of the information systems that support it. states receive thousands of recommendations per year from UN human rights mechanisms, regional bodies, and other processes. And these recommendations need to be clustered, assigned, follow -ups, reported on, translated into concrete actions. And we will hear a lot about that in the coming minutes. And this is where digital solutions, and this is where our research in the past years has really found it important to focus on, has become increasingly important. Digital solutions are really the way that human rights monitoring these days are overcoming these challenges. And so over the past years, through our work on the emergence of digital human rights tracking tools and databases, we have seen a growing ecosystem of tools used by governments, national human rights institutions, civil society organizations, UN actors as well as academic partners, of course. And these tools help to access information first and foremost. reduce duplication, support reporting, improve access to recommendations, and ideally move us closer to implementation. So that's the idea of this digital transformation that also our sector, albeit a little bit slower than other sectors, is going through. However, this ecosystem is fragmented. Many tools still operate in silos. Data is not always interoperable, the famous interoperability question that everybody mentions these days. Importantly, civil society and human rights defenders do not always have equal access to human rights information and to the digital space that we have here in Europe, for instance, and technical innovation does not automatically translate into better human rights outcomes. And then, after all of this introduction, I'd also like to mention, of course, where AI enters. And this is quite a fascinating step that we're all facing. AI and machine learning can help to process large volumes of information. We talked about the thousands of recommendations earlier on. Governments at times have to be reviewed by multiple treaty bodies in one year, and we all know the challenges that that produces. It's important to identify patterns, right, to cluster recommendations, support semantic search, facilitate multilingual access, and potentially help users identify gaps. And so these tools are growingly really filling some of these challenges, and we believe that this is a major step in terms of usefulness, but then the design needs to be especially careful. And in human rights implementation, of course, we all know context matters. Legal nuance matters. Institutional mandates matter. And so... And the voices of affected communities, of course, matter. And AI can support human judgment, but it should not. replace it. And I think that this is something that we hear throughout the halls of these AI days here in Geneva. So the question for the session is not simply can AI make human rights monitoring faster? Of course, there is an element of rapidity. But the better question is how can digital tools and responsible AI make human rights implementation more effective? Yes. But also more inclusive, transparent, and accountable. This is what we're all trying to work on together. So with that framing, I'm very pleased to introduce our excellent intervenants for today. Our first speaker will be Mr. Roberto Cespedes, Chargé d 'Affaires of the Mission of Costa Rica to the United Nations here in Geneva. And Roberto will bring the perspective of a government user. Costa Rica, just as a small introduction, has recently worked with the National Recommendations Tracking Database, the NRTD. And Roberto will speak to what this means in practice for a state in engaging with recommendation coordination and follow -up. And our second speaker will be Marieve Boyer, Human Rights Officer in the Capacity Building Program of the Office of the High Commission of Human Rights. And Marieve is part of the team responsible for the design and support of the National Recommendations Tracking Database. And Marieve will help us understand the logic behind the tool, its design choices, and how perhaps the OCHR sees its future, including some of the AI functions that I find fascinating and that, yes, we hope that will help governments track the implementation of human rights recommendations in a more efficient but also more... responsible manner. And our third speaker is Mr. Lukasz Osiewicz, Assistant Professor at Adam Mikiewicz University in Poznan, as well as Director for European Affairs of NeuroRise Foundation. He's also working with the Heuridox, which is a close partner of ours as well. And Lukasz will discuss academic projects that he's leading that leverage natural language processing and generative AI to build tools called databases for human rights monitoring, really allowing that access to the databases. And this is a particularly important perspective because academic projects can often test new methods and explore prototypes. And so we're extremely delighted that you're all here with us. And yeah, I'll ass over the floor to Roberto. Thank you so much.
Roberto Cespedes
Thank you, Domenico. Good morning everyone here in the room and those joining online. Thank you for being here with us. Thanks to the Geneva Human Rights Hub, Frederick Norman Foundation and the Office of the High Commissioner for organizing this event. So I can speak a bit about Costa Rica's experience in our recommendations, how we track them, how we monitor and how we try to implement. I'll go back a bit just to tell you that we established in 2011 what we call, it's a very long name but it's like the Inter -Institutional Commission for the Follow -up of Recommendations. It's the NMIRF, National Mechanism for Follow -up. So this commission. This commission meets regularly. In principle, once a month, it's coordinated by the Ministry of Foreign Affairs, and it meets with most of the ministries in government, plus representatives from the judiciary and from parliament, plus the NHRI from our country, which is called the Ombudsman, the Defensoría de los Habitantes. And recently, although it had always existed, but it's now institutionalized, we have something called the Permanent Entity of Consultation, which is that it's two representatives from civil society who have a seat at the table at this commission in every meeting. And so what we do is, from there, every session we go through what's going on in each of our institutions, and check for... what they have been doing in implementation, but also how aware are they of what we are receiving, right? Because human rights recommendations will have been there for a long time, but there is a big challenge always of keeping up to date to what has been recommended, what we need to do in that space, what the government can do, and so it's a constant exercise in this regard. So we didn't have a lot of technical tools to do this until a few years ago. We tried a model that sort of worked, but now with the help of the Office of the High Commissioner, we introduced recently the NRTD in Costa Rica. And this has helped a lot to... actually give a lot of visibility to what has been recommended to us by treaty bodies, by the UPR, by special procedures, and to allow institutions to actually understand where they have these opportunities to implement a recommendation. The NRTD, we find it a very, I'm not promoting it, but I think it's very useful. It's a very flexible tool, and it's very intuitive. So we've had, in that sense, really appreciated the help of the office of the High Commissioner in this sense, because we've had experiences so far with the Ministries of Education and the... the health and our institution for child safety. in terms of what we can do, and they have found that the functionalities of the NRTD is quite useful. So in the end, this is about managing data that has always been there, but it's very difficult to access, at least for countries with limited resources, and also a big turnover in the people who are in charge of this. This is another of the challenges that we've faced constantly, but now we've introduced a constant training cycle for people both at our ministry and at all ministries on the NRTD. And I think this is also key because rotation tends to be quite high in some of these places, and maybe someone is there for one, two years, then they learn how to manage the two, and then... that knowledge is lost. So this idea of constantly training people has been incorporated in how we manage the NRTD. And the value, we've just recently started using it this year, but we already see the value of people inputting data and accessing data, and it's been commented in our meetings. And we are very encouraged that, you know, on the even short and medium term, we're going to see that we're going to be able to detect patterns of how people interpret what they're doing and on their inputs, based on their inputs into the system, and with the use of AI, how can this be improved, streamlined, and connected to other people. And this is something that we're going to be able to do in other parts of the government and civil society. I'll just end my first five minutes, sorry, I'm probably taking too long, mentioning that civil society is a big part of our implementation, and also we are moving ahead for them to be able to access in the future the NRTD. So that we don't only improve transparency, but we have civil society as a partner that builds on implementation and monitoring of the recommendations. So it's not only them observing from a distance, but actually making suggestions and being able to see what every part of the state is doing, and then come in and work together. With the state for our proper implementation. I'll leave it at
Domenico Zipoli
Thank you so much, but I have a lot of questions. It is so important also from our perspective of studying these tools and really grounding the practical reality of the use by governments of these tools because it's all well and good when the frame is well designed, but to understand the value of a tracking tool that is not just in storing recommendations, and in and of itself it's a useful thing at that, but also that of helping institutions organize themselves around the actual implementation patterns. So who is responsible? Just the fact of having a place where the responsible actor is identified really cleans the picture in a government structure. What information is needed? What has already been done? This aspect of pattern detection that automatically allows one to even consider possibilities for early warning because then patterns... often lead to some aspects.
Roberto Cespedes
Yes, maybe I'll comment one. Sorry to comment, but one of the functionalities that we found quite useful is the idea of clustering by topic or by theme so that entities have said, oh, okay, I see that this recommendation would enable us to work together with this or this other institution, right? So it increases communication between them and action eventually.
Domenico Zipoli
And sometimes I think that, sorry, I'll again take the floor. On the aspect of NMRF establishment, so not all member states have these national mechanisms for implementation reporting and follow -up. Sometimes what we say is that one way to establish an NMRF is to digitize your work through tracking tools because regardless of your structure, you then identify the different key implementing actors within that digital interface. So digitalization. helps also to institutionalize the work of different governments. Okay, so we've heard the user perspective. Now to the developer side, the initiator of this wonderful project that is now a few years old, I believe, right? Yes, so what is the NRTD? Maybe the trajectory and some of the plans for the future. Over to you,
Marie Eve Boyer
Maria. Thank you very much, Domenico. Can you hear me? Yes, okay, thank you very much. No, indeed, I think it's important to go backwards and understand why we came up with this idea of developing the National Recommendations Tracking Database. So maybe first, what is the problem that we are trying to solve at OHHR? We are really trying to be client -oriented. So the problem is that, indeed, it is, of course, an opportunity to have international bodies looking at the situation of every country We are trying to be client -oriented. We are trying to be client -oriented. assessing where the problems lie, what can be the solutions, providing guidance. This is, of course, an opportunity. But when you are a state, for example, I just came back from Brazil, and you have basically the Inter -American Commission that provides recommendations to you. You have the Inter -American Court that actually makes decisions that need to be complied with. You have the International Human Rights System with the Human Rights Council, the Universal Periodic Review, committees of experts. You also have special procedures, special reporters that visit the country, and Brazil is really a country that has been visited a lot. Just in the last year, I think there have been five special reporters that visited Brazil. So imagine, you know, with all these visits, a number of, you know, I would say problems that have been identified, recommendations that have been, you know, made. You know, where is the coherence among all of this? And I really think that from a human perspective, it's very difficult. sometimes to really think, okay, is there, I mean, this is a fragmented, it looks like a fragmented system, so is there coherence? And I think that actually the digital space can help bring that coherence or can help show that actually there is coherence. Because indeed, as you said, when you start clustering those recommendations, you see that, you know, all of them go in the same direction. There's no big contradiction. Contradiction really usually is, I would say, yeah, more isolated than anything else. So the problem that we are trying to solve is you have a lot of those recommendations. How can you bring all of this together in one place in order to really look at implementation? So actually what we started doing was, from a global perspective, to develop an online platform that is called the Universal Human Rights Index. And this is an online platform. It's an online platform anyone can access where you have all the, I would say, outcomes of the UN system. that you can search per country, per theme, per group, per sustainable development goal or target. And if you Google it UHRI, Universal Human Rights Index, you get to that online platform. And this was the starting point. But then we realized that actually what states needed and wanted, and again, really from a client orientation perspective, what they wanted was moving from standard setting and from guidance to implementation. And in order to do that, you need a tracking tool that enables you to look at recommendations and see their trajectory all the way to change at the community level, basically, at every citizen or every human level. So how do you do that? And that's how we developed the National Recommendations Tracking Database. And to be very honest, this is not the only tools that exist out there, as Domenico said, there are many of them. For us, it's good because the more, the merrier somehow, because this creates a community of developers who can look at the different tools and see, you know, how to improve the system. Also, of course, receiving feedback from states and from users is really key because that's how you can improve the tool. And we are doing it every day, you know, with the different states that we accompany throughout the process. Right now, we have 20 countries that are really using it, but we have basically 40 more that are, you know, waiting for this tool to be deployed in their country, but we just, you know, are lacking capacity also to accompany them. So, there is a demand. This is for sure. So, that's not a problem of demand. So, what we did was then using all the information that is in this Universal Human Rights Index and for that country that is not interested in the recommendations to other countries. Getting and migrating only the recommendations that pertain to that country. And then we basically developed functionalities where that, you know, country can cluster all the recommendations per theme, per SDG, per group, but also per ministry. Because as you said, looking at, you know, the government perspective, what they need to do is, okay, me as a civil servant, what is on my, you know, table basically to implement? What am I supposed to do? And am I supposed to lead it myself? Are we co -leading it with other ministries or other entities? Am I leading and others are going to complement and work together with us? So this is really this perspective that we try to embed in the NRTD. So clustering, assigning responsibility, because accountability is about that. And then it's about reporting. So it's about those ministries and all those entities, you know, to report on what has been done to implement those recommendations. Those different, the guidance, the direct, the... all the information that has been provided to them. And here I think an important aspect is, of course, data. And you talked about interoperability. It's not just somehow like a big word. It's really very, very concrete. Again, coming back to Brazil because I just came back, I mean, data exists. Not all data. There's definitely a lack as far as disaggregated data is concerned because, you know, the idea is to make the invisible populations visible for sure. But data is there. But there's so much data. How do you process this? And that's where AI comes into play, you know, to play because AI for us, and I think the Secretary General said it at the opening, You know, the idea is really to keep control. For us, that's the position of the United Nations. We do not want AI to take control. We want us to take control, but AI can really help. AI can help scale up, you know, the work, can help in basically spotting all the relevant data because data is there. But how do you spot? What is the data that is necessary for us to measure progress? Is there additional data that needs to be collected? All of this is really enabled by AI for us. So we provide AI assistance in the digital tracking tool for the government to look at the recommendations they need to follow up, the data that they have to see what can be used to report on progress, what needs to be actually, you know, developed. And that's really what we are trying to do right now. So that's what I could say. But maybe one thing that is important, I think, when we talk about human components, it's not just a word and saying, oh, yeah, human component. It's essential. You mentioned that some states first develop tracking tools and then build institutions. We are not of the view that this works in the long run. Why? Because precisely, Roberto did not want to mention other tools, but some tools have really shown limitations because there was no human component behind it to feed information in the system. I mean, AI is not going to do that. It's not going to create a reality. The reality is there. It affects people every day. So you need people who are working with communities who know the subject matters and who are able to enter data in the system, but not only enter, analyze it, have this reflection on what works, what doesn't work. What works in the country may not work. What works in the system may not work. What works in the country may not work. And this, for the time being, t least AI is not able to come into play. So getting really a sense of what AI can do and what humans will always need to do behind those tools. Thanks.
Domenico Zipoli
Excellent. Thank you so much, Marie -Eve, for this very important, I'd say, both institutional and design perspective. And I'd say that what you say also reminds us that, in a way, digital tools are not neutral containers. And so the design of digital tools from the get -go may eventually, down the line, also influence how recommendations are being implemented. So the attention on the design of these tools has to be absolutely ironclad because eventually it is how implementation will eventually become. We're going from, perhaps I'm exaggerating, but sometimes. When we discuss these issues, I tend to say that we're going from a... what we studied, right, during our masters on human rights about periodic reporting. I think that in the next five to ten years, now of course we'll head to the academic side of the conversation, but periodic reporting might eventually end up being a term that would be less and less used because it would be a continuous reporting. You know, information will always be there as long as it's open access. And I think this is a fascinating term for In Earth Space, and yeah, and it's a pleasure to live this trajectory with our partners both here on the panel and in the room. Okay. So we move now from existing institutional tools to, and I'm a big, big fan of what you
Lukasz Szoszkiewicz
you do, Lukasz, the academic and experimental space where natural language processing and generative AI are being tested in increasingly practical ways. But I'll leave that to Lukasz to discuss in further detail. Over to you. Thank you. Could you just provide me with the possibility to share my screen? I just sent the request, I think. Okay. So, just one second. Let's see if we've got... Or if you can share the screen, you can just open the link that I shared in chat, and that will be fine. Perfect. Thanks a lot, colleagues. Okay, so thank you so much for the introduction and for very interesting discussions earlier in the panel, and I will adjust my presentation to what I heard because I will build on different bits that you mentioned. Starting from Domenico, I'm representing academician perspective, so I perceive my role as an experimentalist and handing over results of what I do to other stakeholders, where of course the risks should be properly considered before these tools are deployed at scale. And secondly, I think that the Universal Human Rights Index is my favorite tool, and it was inspired by the work of the University of California at the University of California, for many of the things that you can see now. I am showing four of my projects that relate to human rights. You can scan whichever QR code you can see and you will be redirected to one of the tools. The first one, the tool that was the beginning of my journey with natural language processing, was a database of general comments that was subsequently extended to jurisprudence of UN treaty bodies and also thematic reports of UN special procedures. Then the dashboard for the analysis of Universal Human Rights Index recommendations because what I was missing as an academician was the lack of analytical layer. You have great data, you have very granular data, but you don't have plots. You cannot generate and track quantitative trends in data and if you have approximately 300 ,000 records, that's the most efficient way actually to digest what's there in the database. And... In other examples, For example, it's a European Court of Human Rights database that I'm currently developing, because again, here, for me as an academician, it's a great tool, but it can be upgraded, for instance, to allow for a paragraph -level search exactly like Universal Human Rights Index, but on judgments from the European Court of Human Rights. Because when you type in a keyword and you get 200, 1 ,000 results and you have to open all the tabs, it's not very efficient, and we can make it much more efficient today. And here is my first point, that software is just a tool for a task, and it refers to also what we mentioned before, that it used to take a lot of time and a lot of resources, and we had to adjust our needs to the software that was on the market, and to the licenses that we were able to secure. And currently with generative AI, it's exactly the opposite. We have that issue to solve, and with generative AI, we can build the software exactly to our needs, and just show you an example of Universal Human Rights Index. This is one of the applications that I developed. It's bringing together all... data from the Universal Human Rights Index, which is available for download in machine -readable format. And you can see different things like, you know, top countries that have the most records in the database, top themes, all the metadata basically that exists. You can, of course, go to like country profiles and check how, for instance, Switzerland is dealing with Universal Human Rights Index if it comes to themes, themes recommending, most frequent recommending bodies, concerned groups mentioned. So these are all metadata from Universal Human Rights Index, no generative AI involved at this stage. And this is also the important thing that you can leverage AI tools to generate code, scripts that work deterministically, not probabilistically like AI. So instead of uploading data sets to language models, to browser versions, to a version of ChatGPT or any other tool, you can generate scripts that will work exactly the same every time. And then you can... deploy them online like I did, and you can use them and be sure that they will work exactly the same every time. There's no room for hallucination, basically. So instead of using LLM to analyze data, you use LLMs to build tools to analyze data. And these tools, the code can be handed over to IT professionals that inspect it for cybersecurity issues before deploying, privacy issues, and whatever other issues with code can happen. Another example here is General Commons database. It was originally the General Commons database, but it expanded to jurisprudence of treaty bodies, special procedures. You can see that you can, yeah, like, you know, just search any term, basically, using different logical operators. You get the paragraph level, and that's also a good thing about working with UN documents that are very nicely structured. So paragraphs are always preceded by the number dot. It's very easy to extract text from PDFs. And as a lawyer, you typically... I'm a lawyer, and that's important. so caveat that I'm not an IT professional. I met a PhD in law, and I just learned things on the fly, and this will be the end of my presentation because I think it's an important thing to leave that space for experimentation. And here you can see, for instance, that after clicking on paragraph, you get the neighborhood, so like the preceding and subsequent paragraphs of the tool, which is important for a lawyer to understand the context better because the fact that you have a hit in a paragraph does not mean that it relates to the issue that you are trying to research, so we need that context. And this is an example of whatever feature I will think about, I will just communicate in natural language to the most cutting -edge AI model, and it will generate the code that I can host then online. And one of the examples that you can access was built last evening, and you can see what you can do in like two or three hours. And that's the last QR code. This leads me to my second point, and there is a great plot showing how software development changed already. You have a burger that is on the right representing traditional software development. You have the decision layer, so we as stakeholders decide what we need, what kind of disaggregation we need, whether we need paragraph level, document level, what kind of data from which ministries, institutions we need to fetch. Then we have the execution layer. These were three big burgers before, so the code generation. We were handing over technical specification to IT team that was developing the code. We were waiting, sometimes very long, sometimes it was very costly. And then we had the delivery bottom layer, which was basically testing whether it really does what it's supposed to do, doing some tweaks here and back with IT department and then communicating outside. And it's changing now, as you can see. The execution, code generation. It's extremely reliable, extremely easy with generative AI. And if you talk to IT. professionals, most of them will generate most of their code, even sometimes all of the code with AI because it's so reliable. And we can invest much of our time in decision -making layers. So we decide really what do we need. We do different prototypes, which you can see is the result of like 20 different prototypes that I tested. So I develop the full application instead of like a mock -up that is not clickable. So and then I test user interface and I can really arrive with a solution that is doing exactly what I want. And again, the delivery, it's an important thing that especially as a domain expertise and this is something important thing, the design and delivery layer cannot be done by IT professionals. So IT professional that does not human rights law, is not a lawyer, will not know what do you need as a lawyer. I have like 10 years of working with law, so I know that I need paragraph level, for instance. And I don't need the document level, even though maybe it's easier to implement and more efficient from the... performance point of view. And this is something that is to be done by domain experts, so I think that the space in the software development for domain experts like human rights experts is expanding, and the role of IT is decreasing relatively in the whole process. And also if it comes to assessing whether tools work as they should, the most important thing that is now in the context of AI is how to develop benchmarks. So how to measure whether something works as it is intended to work. And this is again something that can be done only by domain experts. It cannot be done by IT professionals who do not understand the human rights context. And to wrap up, this is just the last slide. I arrived to all these different tools by basically experimenting with things. Things that were, when I was starting back in 2022, starting with the launch of JetGPT 3 .5, I think it was like many people thought, well, I don't know how that it was impossible. you just go there, try it, and many other things are possible. And I think it resembles the RPG games. You have the Heroes 4, my favorite game from my childhood, where you cannot see the map, the full map. You just can see the surroundings. So that's the new model. And you think that, you know, maybe I will go this direction, it will work. Sometimes it does not, but sometimes it does. And very frequently it does, and I perceive my role as an academician, as a character that is exploring this territory. And, yeah, I hope that some of the things that I develop can be then used by institutions n international and national levels.
Domenico Zipoli
Thank you very much, Lucas. Honestly, you're such a game changer. And had we not have, if we hadn't this 45 -minute rule, I would have, actually, we did have it and I didn't interrupt it, because I think that this kind of work is super important, really seeing how technology enables the work of a human rights professional institution. It's such an easy, well, not easy, but easier than before. manner perhaps I'm simplifying but.
Lukasz Szoszkiewicz
I think it's easy and that's the point I think that also the mindset of IT developers is that you know the cold crushes very frequently and they are used to it and then just repeat and work and what I experienced because I'm teaching students that law students are not used to it for instance you know something crushes I'm not good at it it's not for me and
Domenico Zipoli
the mindset changing the mindset is very important excellent very good so we do have only five minutes there within or six to be precise it would be yes the time for for some questions and answers I know that we have representatives of tool well teams that have developed great and fascinating tools in the room as well so of course I'm looking at you first so there is question from the online our online colleagues as well yes please
Cecilia de Armas
Yes, hello. Yes, so I was saying that we are the team from the Global Torture Index at the OMCP, so we just had last week or two weeks ago the second edition of the launch. And a lot of the things discussed now, it really touches us in the sense that all the challenges and the good practices and the ups and downs appear on gathering information, gathering data where many states specifically in terms of torture is quite on purpose many times. And then so how to tackle the problem of data accessibility by governments, and then how do civil society enters, right? So now we have working with 100 NGOs at the national level to implement it in 39 countries, so the idea is next year to keep scaling, but then also the challenge of having a lot of data every year, right? So then from one side, we have a lot of data. Then we need to identify which is the data relevant to identify the risk on torture. And then also how do we visualize this and how do we communicate this also to the public because it is not only for us that we are interested in the topic and that we work on it, but then also that it reaches to the communities and then to the people in general. So first of all, huge thanks for this discussion. I think it was great from all the different perspectives. And then in this, we need more governments involved in this, basically. So we need governments to release data, also in terms of the catch recommendations on torture. So this is a huge gap. A lot of human rights violations. We are missing data from civil society to track progress. And then also we are welcome to be in discussions. I think the only, yes, we think that the only way is just to have dialogues with government, with civil society, with academia. And then just to identify the hot topics, let's say, to... to track progress. Because it's true that some governments... are improving things, but then some of these things are not seen or are not visible to everyone. And we also want to avoid this naming and shaming, right? So also to identify the good
Domenico Zipoli
Thank you very much, Cecilia. Yes, please. And if we can be short so that the questions will then also represent, or the replies to the questions will represent your closing thoughts. Thank you very much.
Axel Leblois
I'm Axel Lebreu from G -State. We have been monitoring the CRPD, the Convention on the Rights of Persuasive Disabilities, for the past 20 years. Our current main tool is the DARE Index, the Digital Accessibility Rights Evaluation Index. We work with 143 countries where we have teams of local advocates feeding the data. And my question is very clear, simple, is we structure our research always in structure. We do it in a way that we can do it in a way that we can do it in a way that we can do it country commitments, country capacity, and outcome. And the variables are classified under those three legs. And the structure, process, and outcomes is kind of traditional human rights monitoring structure. And my question is the following. With AI, what we notice is it's pretty simple at this time to actually gather data for the first two legs. You can easily find, so that's research. What's going on in a country for commitments, even processes. But in the outcome field, you are entering a huge field of hallucination without user feedback, without the end user analysis and documented feedback as to what's going on really for the people who are supposed to be protected. So my question to you is how can you get the valid data? For that third leg. And knowing as well that the human rights treaties, bodies at OHCHR have a very long cycle, like several years for the Council to come back nd everything, when actually things are going so very fast in the field.
Domenico Zipoli
Thank you very much. And this also goes hand in hand with one of the questions from our online participants, really also the evolution from monitoring to implementation of this day and age. So we have literally 30 seconds each. So a closing statement, I know I'm well aware of the problem. Oh, five more. Okay, thank you so much. Okay, so we've got five minutes left. But I may also... One more question. I would also like to add that we could also continue our conversations, and I can see that there's a need for that at our booth. It's up at the AI for Good Expo, booth 117. It's close to the Friedrich Naumann Stiftungs booth, so you're more than welcome to join us here. But, yes, I would... I would say if we could start from Lukasz, Marijew, and Robert.
Lukasz Szoszkiewicz
Thank you. And regarding the question on data accessibility, this is... and the government, this is exactly the bottleneck. Because once you have the data, even if it's unstructured with AI, it's very easy to make it structured. But we need data. And if we don't have data, it leads me to the second question. We're coming to hallucinations. So there are ways to reduce hallucinations, though, a lot of them. And there are great studies that are testing different AI models, which are targeting legal domain and non -legal domain. And the reduction in hallucination rate and also in how persuasive they are, it's huge. The difference is huge. So it does not mean... And the tools that I built, for instance, I use large language models for semantic similarity, but they invoke paragraphs verbatim, do not change. So everything is like a decision of a designer, and you can reduce hallucinations to a very, very low level. I wouldn't say that you can entirely get rid of them, but you can reduce them to a very, very low level. But it's... t's a long process, I would say, and it's very tricky still to do it. And we need data for that.
Marie Eve Boyer
Thanks very much, and thanks for your interest and also for the interesting tools that you also have, which is really great. Maybe to just talk about, especially I thought it was really interesting what you mentioned about torture, because right now at OHHR, we're doing a study on all of the information that is being requested by treaty bodies. Because, you know, it's a lot. If you look at the recommendations in the list of issues prior to reporting, so before states report, and then when they receive guidance, there are a lot of actually indicators, and there's a lot that is being asked to states. And this is, we've not talked a lot about that, but, you know, small islands and least developed countries, you know, there's a huge data gap and capacity to collect data. So they need to prioritize. And, of course, here there's also a political aspect. to it, right? So we basically started this research with recommendations from the Committee Against Torture, because here what we saw indeed is that what states need also is some guidance on really the minimum sets of data that they need to collect. And here we have a big problem, because indeed we are already not there in all countries. If you look at prosecutions, for example, and especially on some crimes, gender -based violence, all these aspects are really undercovered. So it's about right now what we're doing is really looking at what is the minimum set of data that needs to be basically collected, and indeed then you need independent institutions to collect that data in order not to be, I would say, hijacked. But I would say that there's also, I think, a lot of evidence that says that the minimum set of data is actually the minimum set of data that needs to be collected. So I think that's a really important point. Thank you. you know, a positive or an optimistic side of things. You were talking about outcome indicators. If you look at the SDGs, SDG indicators are all about outcomes. So the idea is not, you know, it's also to see how we can ork together with the SDGs so that we can, you know, get to this minimum data set that is needed to track progress.
Roberto Cespedes
Thanks. Thank you. And very quickly, the question you posed is quite challenging, but I think it goes back to accessibility of data. And it is a big step for governments to know what to ask, what to record, and then the balance between objectivity and subjectivity of success in implementation is a tricky issue. But I think with these kinds of tools, and especially with, what Lucas was saying, of analysis, this might be able to, in the future, probably very near future, give us a sense of how well we are doing. I think that there is a lot of things that we do as states that go underreported because we don't know that we are doing that. It would be easy for us to say, okay, we input this data, tell some code or program to say, well, this data came in, do you think that this was generated by an action of government and is this complying with a recommendation? Probably the answer is yes, but a lot of people don't know that they are doing it. And then from there, we can move to other models at the community level. to see what's going on. I think constantly asking people what they think is key is something that we need to improve on. And, of course, there's a lot of models on how to properly do this, but that's another opportunity out there. I think it's going to to be probably easier and cheaper to start monitoring satisfaction and then linking it up with what we are doing on the ground.
Domenico Zipoli
Wonderful. Thank you so very much to, above all, our intervenants, but also to everybody here in the room, to all of you online. It's been an absolute pleasure to spend these 51 minutes with you. Thank you for the additional time. And lastly, a big thanks, of course, to our friends from the FNF Human Rights Hub and OHCHR for the discussion. You're more than welcome to join us up at the exposition area for the continuation of this discussion. Thanks a lot. Thank you.

Avertissement : Il ne s'agit pas d'un compte rendu officiel de la session. DiploAI génère ces ressources à partir d'enregistrements audiovisuels ; elles sont présentées telles quelles, y compris d'éventuelles erreurs. En raison de contraintes logistiques (audio/vidéo ou transcriptions), les noms peuvent être mal orthographiés. Nous nous efforçons d'être aussi précis que possible.